Case Attraction Phenomena in German. Markus Bader & Michael Meng. Institute of German Linguistics. University of Jena

Transcription

1 1 Phenomena in German Markus Bader & Michael Meng Institute of German Linguistics University of Jena Running Head: Correspondence to: Markus Bader Institut für Germanistische Sprachwissenschaft Universität Jena Fürstengraben 30 D Jena / Germany (49) (phone) (49) (fax) xmk@rz.uni-jena.de ( )

2 2 Abstract In this paper we investigate how case features are represented and processed during language comprehension. To this end, we explored the phenomena of case attraction in German in a series of four experiments using a speeded grammaticality judgment procedure. Case attraction leads to errors that are akin to subject-verb agreement errors intensively studied in English and other languages (e.g. Bock & Miller, 1991): Under certain circumstances, the case feature of a nominal phrase is erroneously replaced by the case feature of an adjacent relative pronoun with the result that the clause as a whole is perceived to contain a case error. In our experiments, we varied whether the head noun was morphologically case ambiguous or not and whether the case on the relative pronoun was a marked case or an unmarked one. The results revealed the following: (1) Case attraction errors occur independent of whether the head noun is morphologically compatible with the case supplied by the relative pronoun or not. (2) Marked case features can overwrite unmarked case features but not the other way round, where markedness is defined with respect to the syntactic case patterns found in the language. These results, which are similar to what has been found for subject-verb agreement errors, suggest that the syntactic representations computed by the human sentence processing mechanism reflect the syntactic markedness distinctions within the case system but does not encode information about morphological ambiguity. Keywords Sentence Comprehension, Parsing, Case, Attraction Errors

3 3 Introduction Current research into the Human Sentence Processing Mechanism (HSPM) has led to numerous insights about the computation of phrase-structure representations during sentence comprehension (cf. Mitchell, 1994, for a recent overview). However, the process of syntactic analysis cannot stop with a pure phrase-structural representation. Such a representation is not complete as long as it does not contain a certain distribution of syntactic features. In English, for example, subject and verb have to agree in their specification of number features. Therefore, the HSPM does not only have to attach subject and verb to their respective positions within the ongoing phrase-structure tree, but also has to check whether subject and verb agree with respect to number. In this paper, we will be concerned with a further type of syntactic features, namely case features. While these play only a minor role in English (but see Traxler & Pickering, 1996), they are of special importance in languages which have a more elaborate case system than English and which, at the same time, rely on morphological case markings in order to identify syntactic functions like subject or object. In German, the language under investigation in this paper, the position of subject and object is much less constrained than in English. Therefore, the correct analysis of a sentence often depends on case morphology. For example, the fact that in sentence (1a) the first NP is the subject and the second NP the object whereas the reverse holds in sentence (1b) can only be determined by taking case morphology into account. (1) a. Der Professor hat den Lehrer besucht. The-NOM professor has the-acc teacher visited "The professor visited the teacher."

4 4 b. Den Professor hat der Lehrer besucht The-ACC professor has the-nom teacher visited "The teacher visited the professor." Despite its importance for successful language comprehension, the representation and processing of case has not played an important role in investigations of the HSPM. Most psycholinguistic research on German has considered case only insofar as case-morphology provides a means to create unambiguous control sentences in experiments investigating the processing of ambiguous sentences (cf. Hemforth & Konieczny (in press) for a recent overview of research on parsing German). Only recently has the question as to what role case features might play during language comprehension become a research topic of its own (cf. e.g. Hopf, Bayer, Bader & Meng, 1998; Meng & Bader, in press; Schlesewsky, 1996). With respect to the representation and processing of number features, so-called attraction errors have provided a main source of evidence (cf. Bock & Cutting, 1992; Bock & Miller, 1991; Eberhard, 1997; Nicol, 1995; cf. Nicol, Forster & Veres (1997) for an application to language comprehension). An example of an attraction error is provided in (2a) (taken from Bock & Cutting, 1992). In this example, verb and subject do not agree in their number specifications. Instead, the verb seems to erroneously agree with the embedded NP the history books. An important finding with respect to number attraction errors has been that there exists an asymmetry between singular and plural. Whereas errors as in (2a) are produced with some frequency, the reverse kind of error, which is shown in (2b), rarely occurs. (2) a. *The editor of the history books are... b. *The editors of the history book is...

5 5 According to Nicol (1995) (cf. Eberhard, 1997), attraction errors as in (2) result from the plural feature on history books percolating to the noun editor, thereby turning the subject NP into a plural NP. When this has happened, plural marking on the verb is an automatic consequence. The further finding of an asymmetry between singular and plural number, as witnessed by the more frequent occurrence of errors like (2a) in comparison to errors like (2b) - has been traced back to the fact that plural is the marked category within the English number system whereas singular is the unmarked, default number specification. According to Eberhard (1997), plural as the marked category is represented by a specific plural feature - a plural flag - within the syntactic representation. It is this plural flag which is erroneously attracted in errors like (2). Singular, in contrast, is the default number specification and lacks a featural representation. Therefore it cannot be attracted and errors as in (2b) rarely occur. In German, a further kind of attraction error exists: errors which occur during language comprehension and which involve case features (cf. Bader, 1994, 1997; Schlewsesky, 1996). As will be described in detail below, under certain circumstances the case feature of a NP seems to be overwritten by the case feature of an adjacent relative pronoun. When this happens, a sentence which is completely grammatical according to the grammar of German will erroneously be perceived as ungrammatical. Since it looks as if the NP has attracted the case feature of the relative pronoun, we have termed this phenomenon CASE ATTRACTION. In this paper, we will report four experiments that have used the phenomenon of caseattraction in order to further our understanding of the representation and processing of case. These experiments will focus on the following two questions. First, how independent are case features within the ongoing phrase-structure representation computed by the human sentence processing mechanism? In particular, is the phenomenon that we have termed case attraction really due to some kind of erroneous feature migration within the current partial

6 6 phrase marker (CPPM), or is it instead a further instance of a syntactic ambiguity resolution strategy which does not necessarily refer to case features? If the former should turn out, then we would have strong evidence that case features are at least to a certain extent independent from the phrase-structural representation they are associated with, because only such an independence will allow case features to percolate within the phrase-marker. The second question concerns the notion of markedness. As we will describe below, dative case has a marked status within the German case system. Given the fact that the difference in markedness between singular and plural number in English has such a strong effect on subject-verb-agreement attraction errors, one might expect that a similar difference turns up in attraction errors involving case. A main task of the following experiments will therefore be to determine whether case features can be attracted across the board, or whether case attraction is confined to dative case, i.e. to the marked case within the German case system. In the next two sections, we will first give a short overview of case within the German grammar and then introduce the phenomenon of case attraction and the question raised by this phenomenon. After that, four experiments will be presented that have investigated the representation and processing of case. Our overall conclusion will be that case attraction is a genuine process of erroneous feature migration which only affects marked case features. Case in German German has four cases: nominative, accusative, dative, and genitive case. Of these four, only the first three are used productively for arguments of verbs, whereas the main function of genitive case is to mark modifiers and complements of nouns (comparable to English of). We will therefore be mainly concerned with nominative, accusative and dative case in the

7 7 following. As we will show, dative case in German is different from nominative and accusative case with respect to morphological, syntactic, and semantic properties. Syntactically, nominative and accusative are so-called STRUCTURAL CASES which are assigned in specific phrase-structural configurations. Dative case, in contrast, is a LEXICAL CASE which needs a specific lexical licenser, that is, a particular verb which assigns dative case as a lexical property. One syntactic construction where the divergent behavior of structural and lexical case shows up is the passive (cf. (3)). When a sentence with an accusative object is passivized, the object of the active clause becomes the subject of the passive clause (cf. (3a)). If, however, a sentence is passivized which contains a dative object, the object does not turn into a subject. Instead, it still shows up with dative marking, with the consequence that the sentence does not contain a subject but only an object (cf. (3b)). In sum, nominative and accusative case interchange in the active-passive alternation whereas dative case is inert. This inertness of dative case is one of the reasons to classify dative case as a lexical case, that is, a case that is tightly bound to particular verbs. Nominative and accusative case, in contrast, are structural cases that are assigned in certain structural configurations. (3) a. Sie nom unterstützt ihn acc. Er nom wird von ihr unterstützt. She supports him He is by her supported "He is supported by her." b. Sie nom hilft ihm dat Ihm dat wird von ihr geholfen. She helps him Him is by her helped. "He is helped by her." Dative case differs from both nominative and accusative also semantically. Nominative and accusative are compatible with almost any thematic role. Although there are certain

8 8 generalizations of the form If an argument bears thematic role X, it will be realized as a subject and therefore receive nominative case, generalizations in the reverse direction do not exist. If, for example, an NP bears nominative case, this gives no clue whatsoever as to the thematic role of this NP. Dative case, in contrast, is more restricted with respect to the assignment of thematic roles. Somewhat loosely speaking, one can say that dative case is bound to the thematic roles of recipient and beneficiary (cf. Wegener, 1985, for extensive discussion). Let us finally turn to the way case is spelled out in the morphology. Morphologically, dative case is special in two ways. First, dative case is unambiguously signaled most of the time, whereas nominative and accusative case often coincide. This is shown for relative pronouns in Table 1. For definite determiners, the inflection paradigm would be identical to that shown in Table 1 with the exception of the genitive forms. Insert Table 1 about here A second morphological difference between dative case on the one hand and nominative and accusative case on the other hand shows up when we consider the licensing of different arguments by morphological means. Whereas the assignment of dative case to a verbal NP-argument must be licensed morphologically on the NP, the assignment of accusative does not need such a morphological license (cf. Gallmann, 1996; Vogel & Steinbach, 1995). For example, the quantifier viele (a lot) has both an inflected form and a form without inflection. In a position where structural case is assigned, both the inflected and the uninflected form can be used. This is shown in (4a) for an accusative marked object. The same also holds for nominative marked NPs. As shown in (4b), when used as a dative object, only the inflected form is useable. Without a dative inflection on viel, an

9 9 ungrammatical sentence results. This shows that dative case is in need of a morphological spell-out whereas nominative and accusative case are not. (4) a. Peter unterstützt vieles / viel Peter supports a lot (of institutions, campaigns,...) b. Peter widerspricht vielem / *viel Peter objects-to a lot (of institutions, campaigns,...) Our short survey of the case system of German can be summarized by the following statement. Morphologically, syntactically and semantically, dative case is the marked option whereas the two structural cases nominative and accusative are unmarked, default options. Based both on these grammatical generalizations and on an experimental investigation of an ambiguity between accusative and dative objects, Bader et al. (1996) have proposed a syntactic representation of case that gives the marked status of dative case an explicit representation. According to Bader et al. (1996), dative case - but not nominative or accusative case - is associated with a lexical case flag. This proposal can be graphically represented as in Figure 1. The left side of Figure 1 shows the phrase-structure tree of a NP with structural case, here with accusative case; the left side of Figure 1 is the analogous tree for a NP bearing dative case. Insert Figure 1 about here The dative case flag shown in Figure 1 is to be understood in analogy to the number flag proposed by Eberhard (1997) in order to account for number attraction errors in English. It captures the generalization that arguments bearing dative case have to obey stricter morphological licensing condition in comparison to arguments bearing one of the structural cases nominative or accusative (cf. Vogel & Steinbach (1995) and Bayer, Bader

10 10 & Meng (1999) for a more thorough linguistic discussion of morpho-syntactic differences between structural and lexical case in German). The main difference to the representation of number features proposed by Eberhard (1997) lies in the fact that the case flag in Figure 1 is provided not by the noun but by the determiner. This property stems from the fact that case in German is morphologically realized mainly on determiners but only rarely on nouns. A further difference pertains to the fact that there are two unmarked cases nominative for subjects and accusative for objects whereas there is only one unmarked number feature, namely singular. We assume that this further distinction is reflected in the fact that subject and objects are differentiated by virtue of their phrase-structural position as determined by the argument structure of the verb. : The Basic Phenomenon Case-attraction was first reported and experimentally investigated in sentences exhibiting an ambiguity between accusative and dative object (cf. Bader, 1994; Schlesewsky, 1996). In the experiments to follow, we will investigate case attraction using the particular kind of subject-object ambiguity seen in embedded clauses like (5a) and (5b). (5) a. (Ich glaube,) daß Fritz das Buch geschickt hat. I believe that Fritz a book sent has "I believe that Fritz sent the book" b. (Ich glaube,) daß Fritz das Buch geschickt wurde. I believe that Fritz a book sent was "I believe that the book was sent to Fritz" In the active clause (5a), the proper name Fritz is the subject and the following NP das Buch the accusative object of the clause final verb geschickt hat. In the passive clause (5b),

11 11 in contrast, Fritz is the dative object and das Buch the subject of the verb geschickt wurde. Since a proper name is ambiguous between nominative, accusative, and dative case and a neuter NP like das Buch is ambiguous between nominative and accusative case, the sentences in (5) are locally ambiguous with respect to the syntactic functions that the two NPs play. Before the clause final auxiliary is encountered, the sentences are compatible both with a subject-before-object (SO) analysis and an object-before-subject (OS) analysis. The local ambiguity seen in (5) disappears if the proper names are replaced by NPs that morphologically distinguish between nominative and dative case. For example, a masculine definite NP like the uncle appears as der Onkel in the nominative and as dem Onkel in the dative. Both in end-of sentence speeded-grammaticality judgment experiments and experiments using a self-paced reading procedure locally ambiguous sentences as in (5) have been shown to exhibit a preference for the SO-word order, as reflected by a garden-path effect for locally ambiguous OS-sentences (Bader, 1997): Whereas locally ambiguous SO-sentences (5a) did not differ from their unambiguous counterparts, locally ambiguous OS-sentences (5b) are consistently more difficult to process than unambiguous OS-sentences. The finding of an SO-preference in sentence pairs as in (5) is in line with current research on parsing German which has shown an SO-preference in a variety of subject-object ambiguities (cf. Hemforth, Konieczny & Strube, 1993; Friederici, Steinhauer, Mecklinger, & Meyer, 1998; Bader & Meng, 1999). With respect to the source of this preference, we have to note that there is no agreed-upon explanation. In part, this is due to a lack of agreement concerning the correct phrase-structure representation of German sentences. Since this paper is concerned with the nature of case features, nothing of what follows hinges on the particular phrase-structures assigned to sentences as in (5). For the following, we will assume that active and passive sentences of the sort shown in (5) do not differ phrase-structurally from

12 12 each other (cf. Haider, 1993; Reuland & Kosmeijer, 1993) and that the preference for the SO-structure in the absence of a relative clause is due to the Case Preference Principles proposed by Bader et al. (1996). When processing a case ambiguous sentence like (5a) or (5b), a parser abiding by the Case Preference Principles will assign nominative to the first NP due to the joint effect of the two clauses of the CPP. The second NP will then be assigned accusative because nominative can only be assigned once. The net effect will be that the sentence receives an SO-word order prior to encountering the clause final verb. (6) Case Preference Principle a. Prefer structural Case to lexical Case. b. Prefer nominative Case to accusative Case. A further experiment reported in Bader (1997) compared the processing of sentences like (5) to the processing of these same sentences but with a relative clause attached to the ambiguous proper name. Two kind of relative clauses were used. The first type of relative clause contained the relative pronoun der which is unambiguously marked for dative case. An example of an SO-sentence containing this type of relative clause is shown in (7). (7) Ich glaube, daß Maria, der ich gerade begegnet bin, das Buch geschickt hat. I believe that Maria who I just met am a book sent has/was I believe that Maria who I just met sent the book In a speeded-grammaticality judgment experiment, SO-sentences containing a relative clause headed by der showed a sharp drop in the percentages of correct answers. Whereas SO-sentences without a relative clause obtained ca. 79% correct answers, SO-sentences containing a dative-relative clause received only 56% correct answers. Given that fact that the SO-word order is usually strongly preferred in all kinds of subject-object ambiguities in

13 13 German, the finding of a sharp drop in the percentages of correct answers for such sentences is rather remarkable since it indicates that the usual SO-preference reversed to an OSpreference under the influence of a dative relative pronoun. We will call this phenomenon CASE ATTRACTION because it seems as if the dative case feature of the relative pronoun has been attracted by the head noun, turning the first NP into a dative marked NP and the whole clause into a clause with OS-word order. What might cause the usual SO-preference to turn into an OS-preference under the influence of a dative relative pronoun? Two hypothesis have been proposed to account for this reversal. According to the Hypothesis (CAH) (Bader, 1997), case attraction is a phenomenon similar to number attraction. (8) The Hypothesis (CAH) A marked case feature can erroneously migrate within the CPPM. Figure 2 shows what is happening in case attraction errors under the CAH. The left side of Figure 2 shows the partial syntactic structure that corresponds to the head NP and the dative relative pronoun of a sentence like (7). Given the representation of case proposed in Bader et al. (1996), the relative pronoun der bears a dative case flag. Furthermore, the NP as a whole bears abstract nominative case due to the Case Preference Principles. According to the CAH, this dative flag sometimes erroneously percolates up to the head noun Maria, in a way similar to what happens in number attraction errors. If this happens, the whole NP will be specified for dative case, as shown on the left side of Figure 2. The second NP will then be assigned nominative case, again by the CPP, and the net result will be that the sentence exhibits an OS-structure prior to disambiguation. If the sentence then ends with an active verb, as it does in (7), the OS-structure that resulted from case-attraction will be contradicted, leading to the erroneous judgment of the sentence as ungrammatical.

14 14 Insert Figure 2 about here In sum, the CAH claims that case attraction results from a migration of the dative case feature from the relative pronoun to the preceding head noun. This migration of the dative case feature is in no way licensed by the grammar of German, and therefore it represents a kind of malfunctioning on part of the HSPM, in the same sense as number attraction errors as in (2) are due to a malfunctioning of the language production system. An alternative to the CAH is provided by the Parallel Function Hypothesis (PFA) (cf. (9)) which states that the parser deliberately uses case-information on the relative pronoun in order to resolve the ambiguity of the head noun. If the syntactic function of the head noun is locally ambiguous, and an unambiguous relative pronoun is available, the parser will assume that the head-noun has the same syntactic function as the relative pronoun. Under the PFH, case attraction errors do not result from some malfunctioning of the HSPM but from a particular strategy to resolve a local syntactic ambiguity. Various versions of the PFH have been proposed by e.g. Bader (1994), Schlesewsky (1996) and Sauerland (1996). (9) The Parallel Function Hypothesis (PFH): If the syntactic function of the head noun is ambiguous, the parser assumes that the syntactic functions of the head noun and the relative pronoun are identical. While prior experimental research (Bader, 1997; Schlesewsky, 1996) has shown that dative case sometimes is attracted, it is an open question whether the same can happen with nominative or accusative case. This lack of knowledge is caused by the fact that all prior experimental investigations of case attraction only tested ambiguous sentences where the attraction of structural case features would not be detectable given the general parsing preferences in German. For example, the experiment reported in Bader (1997) also

15 15 investigated OS-sentences as in (10). The relative clause in (10) is headed by the relative pronoun die. This relative pronoun, which is compatible with both nominative and accusative case but incompatible with dative case, functioned as a subject in most of the sentences used in Bader (1997). (10) Ich glaube, daß Maria, die mich bald besuchen wird, das Buch geschickt wurde. I believe that Maria who me soon visit will a book sent has/was I believe that the book was sent to Maria who will visit me soon. Sentences as in (10) induced a garden-path effect, but this garden-path effect does not allow the conclusion that nominative case has been attracted since, due to the general SOpreference in German, a garden-path effect is found in such sentences whether a relative clause is present or not. The same is true for the experiments reported in Schlesewsky (1996). Although he used a different kind of syntactic ambiguity, his sentences containing a relative clause headed by a relative pronoun bearing structural case do cause a garden-path effect even in the absence of such a relative clause and therefore do not allow any conclusions with respect to the attraction of structural case. To summarize this section, it has been shown that modifying an ambiguous NP with a relative-clause headed by a dative-marked relative pronoun can induce case attraction. In contrast to dative case, it is an open question whether case attraction can also originate from relative pronouns bearing structural case. The experiments that we will report now will investigate what kind of mechanism causes case attraction and whether case attraction is confined to dative case, as the marked case in German, or not. Experiment 1 The first aim of Experiment 1 was to test between two competing hypothesis about the mechanism(s) responsible for attraction errors, namely the Hypothesis

16 16 (CAH) and the Parallel Function Hypothesis (PFA) introduced in the preceding section. As we have shown above, ambiguous SO-sentences containing a dative relative clause (cf. (11a), repeated from above) are predicted to cause processing difficulties both by the CAH and the PFH. However, a crucial difference emerges if we consider sentences like (11b) where the proper name Maria has been replaced by the definite NP die Frau (the woman). (11) a. Ich glaube, daß Maria, der ich gerade begegnet bin, das Buch geschickt hat. I believe that Maria who I just met am a book sent has I believe that Maria who I just met sent the book b. Ich glaube, daß die Frau, der ich gerade begegnet bin, das Buch geschickt hat. I believe that the woman who I just met am a book sent has I believe that the woman who I just met sent the book Under the CAH, the sentences in (11a) and (11b) should not differ from each other with respect to the frequency of case attraction. To see why, consider the syntactic representations for the head noun and the following relative pronoun given in Figure 3. Insert Figure 3 about here As a comparison of the two trees in Figure 3 shows, (11a) and (11b) have identical distributions of case features under the syntactic assumptions outlined in the preceding section. In both sentences the head NP bears structural case and is therefore without a case flag. The relative pronoun, in contrast, bears dative case and therefore has a case flag. Under the hypothesis that case attraction errors are caused by an erroneous percolation of the dative feature to the head noun, there should be no difference between (11a) and (11b): In both sentences, the dative feature of the relative pronoun der will percolate to the head noun Frau, thereby turning the whole NP into a dative marked NP. Since the determiner die in

17 17 (11b) is as unmarked with respect to case as the nouns Maria and Frau, it should not have any influence on the processing of (11b). (11a) and (11b) are therefore predicted to behave completely identical. A different prediction emerges from the PFH. According to the PFH, the phenomenon of case attraction results from an ambiguity resolution strategy: The HSPM uses the case information on the relative pronoun in order to resolve the ambiguity of the preceding head noun. The proper name in sentences like (11a) is three-way ambiguous: it can bear nominative, accusative or dative case. When encountering a dative relative pronoun, a parser abiding by the PFH will therefore resolve the ambiguity of the proper name in favor of dative case. This in turn will have the effect that the sentence will be assigned an OSstructure which is not compatible with a clause-final active auxiliary. Sentences (11b), in contrast, contains the definite NP die Frau which is only two-way ambiguous, namely between nominative and accusative case. Dative case is morphologically incompatible with such an NP. The presence of a dative-marked relative pronoun should therefore not matter for sentences of this kind, at least under the natural assumption that the parser applies its syntactic ambiguity resolution strategies only in grammatically licit ways. Under this assumption, the parser can resolve the ambiguity of die Frau either in favor of nominative or in favor of accusative, but not in favor of dative case. Sentences containing die Frau modified by a dative relative clause should therefore not differ from the same sentences but without a relative clause. Besides deciding between the CAH and the PFH, Experiment 1 had the further aim of testing whether there is an asymmetry between dative case and structural case, as might be expected given that dative is a marked case within the German case system. To this end, the experiment will also include OS-sentences containing a relative clause headed by a pronoun marked for structural case. While ambiguous OS-sentences (cf. (10)) are expected to elicit a

18 18 garden-path effect given the pervasive SO-preference found in German, unambiguous OSsentences allow to test the marked status of dative case. The crucial comparison concerns the two sentence types shown in (12). In (12a), which is repeated from above (cf. (11b)), the NP die Frau, which bears structural case, is modified by a dative relative clause. In (12b), the case distribution has been reversed; the dative marked NP der Frau is followed by a relative pronoun bearing structural case. Figure 4 shows the parts of the syntactic structures for the sentences in (12) which are relevant to the current discussion (the left tree in Figure 4 is again repeated from above). (12) a. Ich glaube, daß die Frau, der ich gerade begegnet bin, das Buch geschickt hat. I believe that the woman who I just met am a book sent has I believe that the woman who I just met sent the book b. Ich glaube, daß der Frau, die ich gerade getroffen habe, das Buch geschickt wurde. I believe that the woman who I just met am a book sent was I believe that the book was sent to the woman who I just met. Insert Figure 4 about here As already explained, the CAH predicts that case attraction should be observable in sentences like (12a) since the dative feature on der can percolate up to the NP die Frau thereby turning this NP into a dative marked NP. In (12b), in contrast, case attraction should not occur. As shown on the right side of Figure 4, the relative pronoun does not bear any marked case feature and therefore there is nothing that might erroneously percolate up to the NP der Frau. The differing predictions arising from the CAH and the PFH are summarized in Table 2. Table 2 shows how the CAH and the PFH answer the question Does a sentence of this

19 19 particular type induce a garden-path effect?, where garden-path effect is to be understood as showing poorer performance than unambiguous control sentences without any intervening relative clause. For differentiating between the CAH and the PFH, sentences with SO-word-order are crucial. For SO-sentences, the CAH predicts that a garden-path effect is induced by a dative relative pronoun independent of whether the head noun is a proper name or a noun preceded by a determiner. The PFH, in contrast, only predicts a garden-path effect for SO-sentences with proper names but not with definite NPs. Insert Table 2 about here For sentences with OS-word order and the relative pronoun die, the CAH and PFH make the same predictions although for different reasons. The CAH predicts that the relative pronoun die does not have any influence on the processing of the preceding NPs. OS sentences containing such a relative pronoun should therefore behave identical to analogous sentences without a relative clause. Sentences with a proper name are locally ambiguous and should therefore lead to a garden-path effect given the general SO-preference in German whereas sentences with the dative NP der Frau are unambiguous and consequently should not induce a garden-path effect. The PFH also predicts that unambiguous sentences should not induce a garden-path effect. Furthermore, in ambiguous sentences with a dierelative clause the relative pronoun is itself ambiguous and therefore cannot be used to disambiguate the preceding head noun. Therefore, the default-preference for the SO-word order will be invoked for these sentences and a garden-path effect is predicted given that the sentences are disambiguated towards the OS-word order. The sentences described up to this point were all matched pairs of active and passive sentences. A final aim of Experiment 1 was to show that the phenomenon of case attraction

20 20 in no way depends on any peculiarities of the active-passive alternation. To this end, Experiment 1 did not investigate pairs of active-passive sentences but pairs like (13) where both sentences end in an active verb. (13) a. Ich weiß, daß Maria das Buch las. I know that Maria the book read I know that Maria read the book. b. Ich weiß, daß Maria das Buch gefiel. I know that Maria the book pleased. I know that the book pleased Maria. The embedded clause of (13a) contains the verb las (read) which is a simple transitive verb with a subject and an accusative object. (13b), in contrast, contains the verb gefiel (pleased) which belongs to the class of psych-verbs in German which express the experiencer as a dative object and the target-of-emotion as the subject. The sentences in (13) therefore contain the same kind of local ambiguity as the sentences in (5): Until the clause final verb is encountered, these sentences are compatible with both an SO- and an OSanalysis. In (13a), this local ambiguity is resolved towards the SO-structure and in (13b) towards the OS-structure. With respect to their syntactic structures, (13a) and (13b) do not differ from active and passive sentences as in (5a) and (5b), respectively. Method Subjects. 30 students of the University of Jena participated in this experiment. All were native speakers of German and naive with respect to the purpose of the experiment. Subjects received either 5 DM or course credits for participating in the experiment.

21 21 Materials. 30 sentence sextets were constructed according to the two factors word-order (SO vs. OS) and first NP (control vs. ambiguous with relative clause vs. unambiguous with relative clause). A complete stimulus set is shown in Table 3. Insert Table 3 about here All sentence sextets were constructed around pairs of simple SO- and OS-sentences as in (14). These sentences are identical to the sentences in (13) with the single exception that the ambiguous proper name has been replaced by an unambiguous definite NP which always was feminine. Simple sentences as in (14) will serve as control sentences that define the baseline for measuring possible garden-path effects. (14) a. Ich weiß, daß meine Schwester das Buch las. I know that my sister the book read I know that my sister read the book. b. Ich weiß, daß meiner Schwester das Buch gefiel. I know that my sister the book pleased. I know that the book pleased my sister. Unambiguous sentences with relative clause were obtained from control sentences by inserting a relative clause behind the first NP. The first NP in an SO-sentences was always modified by a relative clause headed by the relative pronoun der which is unambiguously marked for dative case (cf. Table 1). Relative clauses inserted in OS-sentences were headed by the relative pronoun die which can either bear nominative or accusative case. In 21 sentence sextets, die received nominative case, in the remaining 9 sextets die received accusative case. The relative clause of an SO-sentence and the relative clause of the corresponding OS-sentence were always matched with respect to the number of words.

22 22 In ambiguous sentences containing a relative clause, a proper name replaced the initial definite NP of unambiguous sentences with relative clause. Otherwise, ambiguous and unambiguous sentences with relative clause were identical. The experimental sentences were divided into 6 sets, each set containing only one version of a sentence and an equal number of sentences in each condition. The resulting lists of experimental sentences were randomly intermixed with 110 filler sentences. A repeated measures design incorporating a Latin square was used. Each subject was exposed to all conditions but saw each experimental sentence only once. Procedure. This experiment - as well as all following experiments - was run using the DMASTER software developed by K. Forster and J. Forster at Monash University and the University of Arizona. Subjects were seated in front of a computer monitor. They were told that they would be presented sentences on the screen and that their task was to judge the grammaticality of each sentence as quickly and accurate as possible. The concept of grammaticality was explained by examples. Subjects initiated each trial by pressing the space-bar. After pressing the space-bar, a fixation point appeared in the center of the screen for 1,050 milliseconds. Thereafter, the sentence appeared on the screen in a word by word fashion with each word appearing at the same position (mid-screen). Each word was presented for 224 milliseconds plus an additional 14 milliseconds for each character to compensate for length effects. There was no interval between words. Immediately after the last word of a sentence, three red question marks appeared on the screen, signaling to the subjects that they now were to made their judgment. Subjects indicated their judgment by pressing the left shift-key for ungrammatical and the right shift-key for grammatical sentences. If subjects did not respond within 2000 ms, a warning line zu langsam (too slow) appeared on the screen and the trial was finished.

23 23 Prior to the experimental session, subjects received practice trials to ensure that they had understood the task. During the practice trials but not during the experimental session subjects received feedback as to the correctness of their judgments. Results Insert Table 4 about here Judgments. Table 4 shows the percentages of correct judgements for Experiment 1. In this and all following experiments, percentages of correct judgments were analyzed with both subjects (F1) and items (F2) as random effects. Two-way ANOVAS (2 word orders 3 first NP) revealed that the factor word-order was insignificant in the subject analysis and marginally significant in the item analysis (F1(2,58) = 1.87, p >.1; F2(2,58) = 3.81, p <.1). The factor first NP was significant both in the subject and the item analysis (F1(2,58) = 4.78, p <.05; F2(2,58) = 4.78, p <.05). The interaction between the two factors word-order and first-np, finally, was highly significant in both analysis (F1(2,58) = 6.23, p <.01; F2(2,58) = 5.77, p >.01). Planned comparisons i were conducted to test the hypothesis derived in the introduction to Experiment 1. Within the category of OS sentences, there were no significant differences between control sentences, ambiguous sentences with relative clause, and unambiguous sentences with relative clause. Within the category of SO-sentences, ambiguous sentences with a relative clause received significantly less correct responses than control sentences (77% vs. 92%; t1(29) = 3.39 p <.01, t2(29) = 3.39 p <.01). Unambiguous sentences with a relative clause received even less correct responses than ambiguous sentences with a relative clause, but this difference was not significant (74% vs. 77%, t1 and t2 < 1).

24 24 Judgment times. Response times for correct judgments are shown in Table 4. In this and all following experiments, response time data were corrected for outliners in the following way. Before analysis of response times, all response times more than 2.5 standard deviations (SD) away from an individual subject's mean were replaced with the cut-off value for the subject (the value equal to 2.5 SD above or below the mean). Less than 5% of the response times were replaced by this criterion. Response times were analyzed with both subjects (F1) and items (F2) as random effects. Two-way ANOVAS (2 word orders 3 first NP) showed that the main effect of wordorder was not significant (both F s <.1) but the main effect of first NP was (F1(2,58) = 11.08, p <.01; F2(2,58) = 4.94, p <.05). The interaction between word-order and first-np failed to reach significance in the subject-analysis but was significant in the item analysis (F1(2,58) = 2.11, p =.131; F2(2,58) = 3.61, p <.05). Planned comparisons revealed the following. Within the condition SO, ambiguous sentences with a relative clause took longer to be judged as grammatical than control sentences but this difference reached significance only in the subject analysis (686ms vs. 605ms; t1(58) = 2.52, p <.05; t2(58) = 1.40, p >.1). The further difference between ambiguous and unambiguous sentences with relative clause was not significant (686ms vs. 697: both t-values < 1). Within the condition OS, unambiguous sentences with a relative clause did not differ significantly from control sentences (647ms vs. 626ms: both t-values < 1) but ambiguous sentences with relative clause received significantly longer reaction times than unambiguous sentences with relative clause (759ms vs. 647ms; t1(58) = 3.47, p <.01; t2(58) = 2.54, p <.05).

25 25 Discussion Experiment 1 has replicated earlier findings (Bader, 1997; Schlesewsky, 1996) that dative relative clauses can induce attraction errors by which an otherwise preferred SOstructure is turned into an OS-structure. Since, in contrast to Bader (1997), we did not use pairs of active and passive sentences but active sentences throughout, we can conclude that the phenomenon of case attraction in no way depends on any peculiarities of the activepassive alternation. In addition to replicating the basic phenomenon of case attraction, Experiment 1 has provided two new results. First, dative relative pronouns induced case attraction errors independent of whether the head noun was compatible with dative case (proper names) or not (definite NPs). This finding was predicted by the Hypothesis but not by the Parallel Function Hypothesis. Second, attraction errors for sentences where the head NP was a definite NP only occurred with dative relative clauses but not with relative clauses headed by a relative pronoun specified for either nominative or accusative case. Whereas ambiguous OS-sentences showed the usual garden-path effect in accordance with the general SO-preference found in German, unambiguous OS-sentences containing a relative clause did not differ from OS-control sentences. Finding case attraction only in sentences containing a dative relative clause supports the assumption that dative case, as the marked option within the German case system, is represented by a case flag within the CPPM whereas nominative or accusative case are not. The fact that we did not find a performance difference between unambiguous OSsentences with and without relative clause shows that the insertion of a relative clause per se did not increase processing difficulty in the speeded grammaticality judgment task. The finding of a reduced accuracy and higher reaction times for SO-sentences containing a dative relative clause in comparison to SO-sentences without a relative clause can therefore be completely attributed to the effect of case attraction. In order to get an estimate of the

26 26 frequency of case attraction, one might subtract the percentages of correct responses for attraction sentences from the percentages of correct responses for control sentences (92% - ca. 75.5%). However, there are two reasons why this gives only a rather crude lower bound of the true percentages of case attraction. First, the possibility of reanalysis has to be taken into account. When the parser detects a case violation on processing the clause final verb in sentences where case attraction has occurred, it might invoke processes which, when completed successfully, will lead to the response grammatical despite the prior occurrence of an attraction error. That reanalysis indeed occurs in the sentences under consideration is suggested by the finding that judging attraction sentences correctly as grammatical took more time than judging corresponding control sentences as grammatical, a finding which is regularly found when garden-path sentences are judged as grammatical. A second reason why simply subtracting the raw percentages gives only a lower bound on the true probability of attraction errors lies in the fact that - even when the possibility of reanalysis is taken into account - one cannot assume a priori that on each trial where dative case has been attracted people will give the response ungrammatical. This assumption would only be licit if corresponding ungrammatical sentences would be judged as ungrammatical with near perfect accuracy. However, for ungrammatical sentences containing case violations, prior research has shown that this assumption is often not true (cf. Meng, 1997; Bader, Meng, & Bayer, in press). To get a better estimate of the frequency of case attraction, Experiment 2 will therefore include ungrammatical sentences which contain the same kind of ungrammaticality as is present when case attraction has happened in sentences like (11). A certain concern one might have with respect to Experiment 1 lies in the fact that the crucial first NP in this experiment was always a feminine NP. For feminine NPs, with the exception of feminine proper names, nominative and accusative case share the same form

27 27 while dative case has a form of its own (for example die for nominative and accusative but der for dative). The two-way ambiguity of die opens up several alternative explanations of the results provided by the current experiment. For example, it might be that the crucial difference between die and der has nothing to do with the opposition of structural and lexical case but with the opposition of ambiguous and unambiguous marking of case instead. Furthermore, given the two-way-ambiguity of die, the HSPM might consider to apply a syntactic ambiguity resolution strategy like the Parallel Function Strategy and in doing so overlook that an NP like die Frau is not compatible with dative case. Experiment 2 will take up these concerns by comparing the processing of feminine NPs to the processing of masculine NPs for which not only dative case but also nominative and accusative case have distinct lexical forms. Experiment 2 Experiment 1 has shown that there is an asymmetry between the two feminine relative pronouns der and die. Whereas the former triggers case attraction errors, the latter does not. We have interpreted this asymmetry in terms of syntactic markedness. Whereas die bears unmarked structural case, either nominative or accusative, der bears marked dative case. However, there is a further asymmetry between die and der. As a feminine relative pronoun, der is unambiguously marked for dative case whereas die is two-way ambiguous between nominative and accusative (cf. Table 1). Therefore, a purely morphological definition of markedness, according to which ambiguous items are unmarked and unambiguous items marked, would be as compatible with the results of Experiment 1 as our definition of markedness in syntactic terms. In order to distinguish between a syntactic definition of markedness (structural versus lexical case) and a morphological definition (ambiguous versus unambiguous form), we

28 28 have to consider masculine NPs. When applied to masculine NPs, the two definitions of markedness give different results. Masculine NPs (besides proper names) have a separate morphological form for each case. Under a morphological definition of markedness, they are therefore all on a par, namely marked. Under a syntactic definition of markedness, in contrast, the same asymmetry holds as for feminine NPs: nominative and accusative NPs are unmarked whereas dative NPs are marked. The syntactic and the morphological definition of markedness therefore make differing predictions when it comes to sentences like (15). These sentences are identical to the sentences in (11) investigated in the preceding experiment with the single exception that the first NP is a masculine NP instead of a feminine one. In (15a), the masculine proper noun Peter is followed by the relative pronoun dem, which is unambiguously marked for dative case. (15b) also contains a dative relative clause which, in this sentence, modifies the definite NP der Mann, which is unambiguously marked for nominative case. (15) a. Ich glaube, daß Peter, dem ich gerade begegnet bin, das Buch geschickt hat. I believe that Peter who I just met am a book sent has I believe that Peter, who I just met, sent the book b. Ich glaube, daß der Mann, dem ich gerade begegnet bin, das Buch geschickt hat. I believe that the man who I just met am a book sent has I believe that the man who I just met sent the book Partial syntactic structures for the sentences in (15) are shown in Figure 5. The top of Figure 5, which shows the representation of a proper name modified by a relative clause, corresponds to (15a) both under a morphological and a syntactic definition of markedness. It does not differ from the syntactic structure for a NP with a feminine proper name (cf. Figure 3). The tree on the bottom left side of Figure 5 corresponds to (15b) under a syntactic